Challenging the assumptions that drive current AI narratives

The public discourse on the future of AI (and our societies’ future with it) is based on assumptions that are not always clear to the casual reader, yet they strongly shape the narratives and their perceived likelihood. In what follows, we treat eight of these assumptions in one section (~a one-pager) with the goal of opening up the space of possible futures by drawing firmly held assumptions a little further back into the realm of the unknown (where assumptions belong).

Illustration of a calm path leading away from current hype

by Thilo Stadelmann and Renate Stadelmann1.

How to use this material

The eight2 assumptions are:

Each one-pager consists of (a) the assumption in its popular (not scholarly) form and (b) arguments to see it in a new light (not necessarily arguments to claim the opposite). Part (a) is deliberately kept short, with a descriptive title (phrased to help classify quickly each one-pager’s main topic), a catchy phrase that formulates the assumption (deliberately hold on a popular level as it might appear in popular news, without scholarly nuance), and a list of few implications for the future of AI and society that hinge on that assumption.

Part (b) contains 3 subsections:

  • Evidence: Arguments and publicly underrepresented background information that challenge the assumption or its status based e.g. on empirical facts, scientific claims, or journalistic arguments.

  • Outlook: Arguments concerning the future for which necessarily no empirical evidence exists, yet, but that follow from the evidence above.

  • Possibilities: Some alternative future developments concerning AI and society that become more likely (even probable) once the popular assumption loses its might; only deliberately incomplete sketches to not preempt the scenario development process.

Each one-pager is kept intentionally brief to serve as a discussion starter. It provides arguments and references, but for brevity’s sake, the arguments are dense and incomplete; the references are meant to provide lacking rigor. Some assumptions are interdependent; if so, this is marked by a linkt to the referenced assumption. Due to the interdependence, the full picture only emerges after a full read of all 8 one-pagers (i.e., some counter arguments, outlooks and opened-up possibilities on related one-pagers enriches one’s understanding of another).

Many assumptions hinge to a certain degree on opinion, hence the evidence often points to a diverging opinion and its foundations, not proof of the opposite assumption. By design, the assumptions represent a certain mainstream view while the challenges appear less mainstream. This is true with respect to the mainstream reporting in popular news and corresponding debate. Agreement to the challenge arguments presented here naturally varies from case to case, but to the best of our knowledge, none represents a niche view, so that both the assumptions and challenges are each represented by (varying) sizeable parts of the experts community.

The most fundamental assumptions that influence many others are on “sameness of human and machine” as well as “imminence of artificial general intelligence” and hence put first. The most direct, practical ones might be “massive job loss” and “de-skilling and loss of agency”, which are presented last for a natural flow from fundamental to tangible. As all 8 one-pagers have cross-references where appropriate or necessary, you can pick your own reading.

Fundamental sameness of human and machine

Assumption #1

“There’s no fundamental difference between a human mind and a machine. Anything a person can do or be, a sufficiently advanced AI system could do or be too.”

Implications

  • AI systems could be conscious, which means that we should not deny them similar legal status as persons (at the same time, human rights on the basis of human specialness become questioned as it surveillance enables the quantification of the extent to which an individual complies with societal norms and hence might deserve such rights)

  • As humans exhibit “general intelligence” (the capability to solve previously unknown problems and deal with completely novel situations with an extremely wide scope), we must assume that machines will have the same powers eventually by virtue of sameness, irrespective of no evidence whatsoever

  • Leads to Artificial general intelligence (AGI) assumption without further technical evidence or known path to implementation

Challenging the assumption

Evidence

  • Today’s AI systems are based on machine learning methods, which manipulate symbols (such as tokens/words) by their statistical shape, not their meaning. Such a system can produce fluent answers about a topic without understanding a word of it (the “Chinese room argument”) [Behavioral and Brain Sciences, 1980].

  • The argument about if this is a difference between pattern matching and comprehension is a difference in degree or in kind is philosophical rather than technical, drawing from one’s stance on anthropology; “computer anthropologies” that argue for “substrate independence” of human properties such as the mind and consciousness are currently common in the tech industry, but highly contested in the humanities [AI and Ethics, 2023; AI & Society, 2021].

  • “Computer anthropologies” tend to be intertwined with other philosophical commitments that in turn shape the thinking of people in the tech industry, specifically in Silicon Valley, to a much higher degree than in the total population [Empire of AI, 2025]. These views underly much of the public discourse through institutional media releases, supported think tanks, high-reach social media influencers, and a network of like-minded researchers [First Monday, 2024a]; and lean into a direction that can appear quite anti-human [First Monday, 2024b].

  • Reports about machine consciousness [Anthropic, July 2026] are labeled misleading and unfounded by peers [G. Seiberth, 2026; Financial Times, 2026; M. Suleyman, 2026]

  • “Neural networks” in AI and computer science (and many of the other terms originally used for humans in this context) are very loose metaphors of things in biology: never meant to be understood as “computer brains”; AI is not so much about intelligence per se but the simulation of intelligent behavior [Science, 2024; AI Magazine, 2006].

Outlook

  • In the absence of any clear scientific evidence in support or falsification of the hypothesis, any stance on this issue is a deliberate choice in worldview, not a technical or scientific matter.

Possibilities

  • AI systems do not need to be conceptualized as necessarily becoming human. The field of AI’s goal is to simulate functions of intelligence in increasing quality and generality, but simulating “a human” is just one way to think about this. Many of today’s narratives hinge at the imagination that AI necessarily will be a kind of computer “being” (including intentions, goals, etc.). What if not?

  • One example in the fiction literature genre is the Star Fleet’s ship computers in Star Trek: powerful AI as mere tools.

Artificial general intelligence (AGI)3

Assumption #2

“AI will soon become as good as, or better than, the single best human at every task that matters economically.”

Implications

  • AI regulation today must be made with imminent AGI and respective “AI safety” (i.e., an interdisciplinary field focused on preventing accidents, misuse, or other harmful consequences arising from AI systems”) in mind

  • AI safety needs strong “alignment” (i.e., methods in the field of AI to steer AI systems toward a person’s or group’s intended goals, preferences, or ethical principles) or otherwise superhuman capabilities will not be controllable

  • Superhuman capabilities will lead to large-scale societal disruptions with potential utopian (“AI accelerationist” view, serving as justification for massive private investments by invoking “humanity’s last invention” and “AI race” narratives; see Accelerated hypercapitalism assumption) or dystopian outcomes (“AI doomer” view, specifically as source of fear behind Massive job loss and Existential threats assumptions)

Challenging the assumption

Evidence

  • AGI played a minor role until the founding of Google DeepMind and OpenAI in the mid-2010s, who adopted the narrative while seeking previously unheard-of investments to fund their research and development; that it is taken seriously has been called “the most consequential conspiracy theory of our time” [MIT Technology Review, 2025].

  • The idea of a technological “singularity” (i.e., the hypothetical event in which technological growth accelerates beyond human control), which is the necessary precursor to AGI, is a (contested) philosophical concept, not a technological inevitability [Journal of Consciousness Studies, 2010; MIT Technology Review, 2011]. For example, it rests on the assumption that ‘intelligence’ is a one-dimensional scale on which consecutive step improvements do necessarily reach a point in which an AI system, applied to its own improvement, can produce the next little step and so forth (“recursive self-improvement”); this assumption is wrong [Wired, 2017; F. Chollet, 2019; Y. LeCun, 2024].

  • Hence, concepts around AGI (including AGI itself and its philosophical underpinnings, see also Fundamental sameness of human and machine) are portrayed as quasi-religious (based on worldview, with a eschatological outlook to the future), not scientific [Oxford University Press, 2010; Cambridge University Press, 2024].

  • Scientifically, there is no known path from where we are technologically to developing AGI. In particular, Large Language Models (LLMs, the foundation of current AI) are largely believed to not leading to considerably higher capabilities than we have today, and suggested next steps are vague (“we need to move closer to the real world”) rather than a clear paradigm to implement [Fortune, 2026].

  • Even though many AI experts today can imagine AGI-like systems in a more distant future based on their worldview, its imminence gets rejected by the vast majority [Forecasting Research Institute, 2026].

Outlook

  • Much of today’s popular narratives about our future under the influence of AI are fueled by the assumption of incomprehensively higher capabilities of such systems than we currently have (see links above).

  • There is a growing conviction also amongst earlier supporters of this assumption that AGI will not be achieved by following the current paradigm of “scaling” LLM-like systems (scaling refers to the simultaneous enlarging of the training data set, the model’s size or capacity to learn, and the compute used to train the model) [Reuters, 2024].

Possibilities

  • How does a future look like in which AI is a very capable tool, but more a consequential further development of what we have today rather than a disruption?

Existential threats

Assumption #3

“A powerful AI that isn’t properly aligned with human values could, by accident or on purpose, cause the extinction of humanity.”

Implications

  • AI regulation must predominantly serve to mitigate such risks, because the likelihood of extinction might be small, but the impact would be disastrous

  • Threats could arise from deliberate misuse (e.g., for bioweapon creation), by accident (e.g., following from a powerful AI system maximizing a trivial objective, thereby destroying everything else in pursuit of a narrow goal: the “paperclip maximizer” thought experiment), or by an artificial general intelligence (AGI) system gone rouge (e.g., the system developing the thought that earth would be better off without humans; see also Artificial general intelligence (AGI) assumption)

  • Competition and open dissemination (e.g., open-source AI) are met with suspicion and massively lobbied against (see also Grave danger of misuse assumption)

Challenging the assumption

Evidence

Outlook

  • See Grave danger of misuse assumption for an outlook on the mitigation of current risks.

  • Indicators suggest that even AI companies otherwise keen to keep fears of existential threats an issue of discussion are gearing their actual risk mitigation efforts into these more tangible directions: OpenAI dissolved their “superalignment” team responsible for long-term risks [CNBC, 2024], and authors from DeepMind, Anthropic, and OpenAI together with academics recently argued for “positive alignment” that should not just focus on safety, but take human flourishing into account [R. Laukkonen et al., 2026].

Possibilities

  • Taking fear for one’s life out of the options coming from having AI opens up new ways of thinking clearly about the technology we actually have and what it could lead to (positively: how does it enable a future worth living in; negatively: how to mitigate actual dangers) and frees resources to do so.

Grave danger of misuse

Assumption #4

“Already the AI systems we have right now are so capable — for example at hacking computers or helping design bioweapons — that they are a threat on the level of nuclear weapons.”

Implications

  • AI must be tightly controlled by governments and never fall into the wrong hands

  • Only a few companies and institutions should be allowed to build and provide AI

  • Open diffusion (e.g., by open source) must be banned

Challenging the assumption

Evidence

  • There is a substantial reality gap between what is shown in isolated benchmarks by researchers and companies (showing huge increases in specific capabilities) and matched reliability benchmarks (testing for consistency, robustness, etc.), where the improvement over the last 2 years was merely ~5-10%; so far, nothing about current systems’ actual real-world behaviour has been unexpected in a specifically catastrophic way the assumption implies [International Conference on Machine Learning, 2026].

  • Large AI companies have a track record of publishing warnings about the capabilities of their models also as a marketing means at least since 2019 [MIT Technology Review, 2019; The Decoder, 2016]. Recent warning of misuse potential mix elements of concern for security experts [S. Willison, 2026] with a staging by the companies that makes the results unsurprising to experts but that is usually left out of the public discourse (“AI escaped the safe environment” and “the model pursues own goals” narratives) [Die Zeit, 2026]; at the same time, they coincide with respective companies’ plans to raise massive capital via IPOs (initial public offering to be listed at a stock market exchange) [TechCrunch, 2026; Forbes, 2026]. See also Artificial general intelligence (AGI) assumption for the spinning of convenient narratives and Existential threats assumption for dangers extending towards extinction risks.

  • The ones mostly profiting from a “dangers of misuse” debate are the currently leading AI companies: Their financial interests would directly profit from any governmental action that makes market entry harder for (future) competitors, e.g., in other regions of the world and through open source. Coincidentally, these companies pay corresponding big lobbying efforts, drive respective mainstream narratives with staged press releases, and support or initiate calls for bans on open source as well as prohibitive regulation in an attempt of “regulatory capture” [Forbes, 2026; Washington Examiner, 2026; The Wall Street Journal, 2024; Benzinga, 2026].

Outlook

  • Looking at real dangers of misuse of today’s AI systems, realistic ways of mitigating these risks have been suggested and partially implemented (e.g., for bioweapons [K. Esvelt, 2022; A. Ng, 2023], cybersecurity [Palo Alto Networks, 2026], misinformation [Internet Pros, 2026], and in general [NIST, 2026]). These frameworks and proposals do not suggest treating AI as a weapon of mass destruction.

Possibilities

  • Separating legitimate safety engineering from policy agendas that conveniently justify market concentration enables open, competitive markets.

  • Open markets, one of the important bedrocks to prevent vendor lock-in and ensure certain technology sovereignty in countries outside the U.S. and China, currently hinges on the availability of open-source AI (so that many specialized and probably local companies can build products around existing foundation models; see Accelerated hypercapitalism assumption).

Accelerated hypercapitalism

Assumption #5

“Powerful AI marks the final stage of the economy. Whichever company or country gets there first will end up owning everything.”

Implications

  • Tech feudalism, in which a few companies become more powerful than even large states, seems inevitable given that digital markets exhibit winner-take-all properties

  • There will be an “AI race”, predominantly between the U.S. and China, with no second thoughts regarding side effects and all other regions of the world becoming mere rule-takers

  • There will be a continued building frenzy for data centers as compute capability is currently the main driver of development and the end here justifies the means

Challenging the assumption

Evidence

  • The notion of a “race” makes most sense based on the assumption that AI is “humanity’s last invention”, i.e., that an imminent powerful AI will overtake the economy, making the one reaching AGI first winner in a winner-takes-it-all game; see Artificial general intelligence (AGI) assumption on the hypothetical nature of this thinking. It is increasingly seen as unrealistic, especially short-term [G. Marcus, 2023; Taipei Times, 2026; Bloomberg, 2026].

  • Hence, because also in the absence of powerful AI, a “race” happens (i.e., the usual competition to be the market leader in still sizeable market, even if not encompassing all of the economy), market leaders seem to have silently changed the strategy [Fortune, 2025]: Instead of betting all on the arrival of powerful AI, resolute geopolitical rhetoric, tough industrial policy [Moore’s Law is Dead, 2025], and attempts for regulatory capture (see Grave danger of misuse assumption) are applied while repurposing the giant build-out of data centers from AGI-producing to vendor lock-in [McKinsey, 2025].

  • Yet, with open-source models catching up increasingly fast, the market is much more open than thought: The capability gap between proprietary frontier systems (from the U.S.) and open-weight AI systems (from China) almost vanished [VentureBeat, 2026] from an earlier 3.3% [Stanford University, 2026].

  • As underlying foundation models thus get commoditized quickly, the application layer in designing AI systems becomes decisive [Andreesen Horowitz, 2023], which is good news for companies also outside the U.S. and China: Advantage comes from capital, proprietary interaction data and talent rather than a unique model architecture.

Outlook

  • An open market with free choice of vendor of agentic AI systems should be one of the most important strategic concerns of every economy: When the application is where the value creation happens, this opens up possibilities for a globally distributed AI industry that caters to local communities and specific sectors, rather than giant generalists.

  • Openness can be assured by public procurement decisions (e.g., favoring more local suppliers to contribute to tech sovereignty) and permissive open-source policy; also enforcing antitrust laws can keep the market from consolidating around a single provider.

Possibilities

  • If legitimate security concerns can be disentangled from regulatory capture attempts by policymakers, markets could safely be considerably more open than today.

  • A multi-polar world with more evenly distributed returns on AI could be the result.

Prohibitive energy consumption

Assumption #6

“AI consumes enormous amounts of data, energy and water — it’s fundamentally unsustainable and bad for the planet and every species living on it.”

Implications

  • AI development in principle is morally questionable

  • Its deployment should be phased out similar to nuclear energy for environmental considerations

Challenging the assumption

Evidence

  • The current AI paradigm is inherently wasteful with respect to data (“sample efficiency”) and compute [MIT Initiative on the Digital Economy, 2020; Association for Computational Linguistics, 2019].

  • Leading researchers argue that this is mainly due to the training algorithm for the underlying “neural networks” (called “backpropagation of errors”): it operates globally on the whole net at once, while for example the brain uses local ‘computations’ [Nature, 2020].

  • The human brain draws ca. 20 watt of continuous power [Journal of cerebral blood flow and metabolism, 2001] and heard in the order of tens of millions of words in total until age 13 [Trends in cognitive science, 2023]. In comparison, GPT-3 (the model on which’s successor the original ChatGPT was built) used 1.287 GWh of power for training [D. Patterson et al., 2021] and was trained on ca. 300 billion “tokens” (sub-word units). According to these numbers, a human brain uses ca. 2.28 MWh up to age 13, which is ca. 320-430 times less energy than GPT-3 until end of training (taking into account that a child’s brain uses more energy, not taking into account that the human brain does more than language acquisition in this time); and estimating a word at 1.3 tokens on average, the human brain also listened to 1500-2250 times less words until age 13 than GPT-3 during training. From this comparison, the current AI paradigm appears >2 orders of magnitude less energy efficient and 3-4 orders of magnitude less sample efficient.

Outlook

  • A next level of AI, not necessarily in capability, but in orders of magnitude higher efficiency, could hence come from more local, neuroscience-inspired training algorithms (and a respective, emerging ecosystem of tools, software frameworks, network architectures, and specialized chips) [Science, 2014; Neural Computation, 2026].

  • Many universities have ramped up activities at the intersection of neuroscience and AI as they see this opportunity [Nature reviews neuroscience, 2025], with researchers naming a 5-year horizon for its realization [NSF workshop report, 2026].

Possibilities

  • AI systems based on a such a “next level” paradigm could easily run locally on one’s smart personal devices, overcoming the energy concern.

  • Huge investments into data centers as a requirement to build or serve AI systems would be unnecessary, enabling a wider distribution of AI vendors throughout the global economy (see also Accelerated hypercapitalism assumption).

  • The collapse of compute as moat and requirement for AI power could lead to market upheavals concerning current investments with an outlook on data center build-out. It could also lead to compute infrastructure becoming public infrastructure similar to streets after a government bailout.

  • Higher sample-efficiency could overcome the current training data bottleneck, leading to better AI systems also for niche tasks apart from text and image analysis and generation, e.g., for specific sectors, industries, and regions.

Massive job loss

Assumption #7

“AI is already destroying thousands of jobs and will cause massive unemployment very soon.”

Implications

  • Now: no way from junior to senior

  • Future: a permanent “useless class”, no dignified occupation for humans

Challenging the assumption

Evidence

  • Junior roles: US data shows no correlation between an occupation’s exposure to AI and a reduction of junior roles between 2022-2024 [Financial Times, 2025]. Currently, a tougher situation for new graduates might be partly due to AI [Stanford Institute for Economic Policy Research, 2026], though the overall decline in entry-level jobs started before ChatGPT and may likely be induced by raised interest rates of federal banks [Economic Innovation Group, 2026].

  • General labor market movement: AI’s effect on the job market is likely small [Stanford Institute for Economic Policy Research, 2026; National Bureau of Economic Research, 2025]. Several layoffs that received high publicity have been identified as “AI washing”, where over-hiring during the pandemic and general cost-cutting was justified by “AI efficiencies” as an excuse [The New York Times, 2026; A. Narayanan & S. Kapoor, 2026]. Klarna, for example, hired people back after replacement backfired in terms of quality and public opinion [Fortune, 2025].

  • Anecdotally, numbers from occupations with high AI exposure show: US software-development job postings grew ~15% in the year after Claude Code’s Feb-2025 launch (mostly in senior roles) while total job postings fell 7% over the same period [Indeed Hiring Lab, 2026]. Employment of technical specialists, people-facing care roles and managers grew between ~5-8% from 2023-2025 while back-office roles declined by ~3.5% [The Economist, 2026]. Radiologists are in high demand in the US [American Journal of Roentgenology, 2025] despite Nobel-prize winning AI researcher Geoffrey Hinton predicting complete AI takeover that capability-wise would even be possible [G. Hinton, 2016].

  • Higher volatility is expected since late 2025 (when agentic systems emerged that can work on long-running tasks than the previous directly responding chatbots). This is might likely influence many roles, inducing changes towards a “Co-Pilot Economy” of human augmentation (rather than replacement) in one scenario [World Economic Forum, 2026].

Outlook

  • Future more extreme labor market breakdown scenarios hinge at the assumption of exponential increases in AI capabilities (towards artificial general intelligence/AGI or superintelligence) [World Economic Forum, 2026]; see Artificial general intelligence (AGI) assumption.

  • Apart from unpredictable breakthroughs, AI’s economic impact lags its capability by decades [Columbia University, 2025]. As AI gives humans more power to change their environment, their work shifts from doing to controlling, from building to deciding what to build in the first place and then evaluating the results [International Conference on Machine Learning, July 2026]. As AI makes predictions cheaper, the price (value) of human judgment rises [International Security, 2022].

Possibilities

  • After 2-4 years of higher volatility, settling could take place: workers adapt workflows and tasks; companies adapt hiring practices (e.g., with respect to longer/new education cycles from junior to senior, in sync with the education system).

  • There might be considerably more jobs in the future as attention shifts to what is not automated, including currently inexistent tasks, roles, and occupations (think of what the development of electricity, the car, computer, or internet brought).

De-skilling and loss of agency

Assumption #8

“Using AI regularly weakens people’s own mental skills (e.g., reading, writing, judgment, agency), leaving them less capable, lonelier, and less prepared for life.”

Implications

  • “AI slop” eats up efficiency gains, leaving every user and every organization relying on it worse off

  • Individuals flee to frictionless artificial relationships, eroding societal cohesion on the micro level (partnership, family, community) further

  • De-learning of unused (hence atrophying) skills is inevitable, and the temptation to use AI tools as a replacement rather than an augmentation of own thinking and judgment becomes irresistible

  • The gap between those augmenting their own skills through proper use of AI and those using it for their own detriment will widen dramatically

Challenging the assumption

Evidence

  • The risk appears real also to those typically downplaying exaggerated AI fears [Public Money & Management, 2026]. However, mitigation strategies are emerging: For example, “pro-human AI design” and “positive alignment” approaches particularly optimize not just for efficiency, user retention or stickiness, but the user’s long-term flourishing including autonomy, growing competence, or healthy relationships [uDay XXIV, 2026; R. Laukkonen et al., 2026].

  • The task of deliberately building the human capacity to use this new level of automation well is multifaceted: Besides AI design efforts from the technical side [Oxford University Press, 2022], moral formation and the furthering of judgment capabilities are tasks for the education sector and parts of civil society that already played this role during earlier transformations (e.g., faith-based communities, associations, neighborhoods) [The Holy See, 2026].

Outlook

  • Naming de-skilling as a specific design target instead of just a worry allows tools, workflows and habits to be built that actively strengthen human competence and connection alongside AI use.

  • AI development is mainly driven by chasing high scores on accepted benchmarks like the LLM Arena. By measuring the pro-humanness of AI systems publicly, the entirety of the field could be steered into to direction of more holistic and pro-human design.

Possibilities

  • Pro-human AI systems are systems designed for being (mentally, psychologically, emotionally) safe for people, and hence more trustworthy with respect to individual safety. Trust can be a decisive economic factor, especially when users get dissatisfied with the current tools.

  • A trust economy could replace, in part, the attention economy (and its incumbents) were the service is not just about entertainment (as with social media), but one’s most inner thoughts and deliberations (as with an agentic AI tool).

Footnotes

Cite as:

Thilo Stadelmann and Renate Stadelmann. “Challenging assumptions that drive current AI narratives”. In: AIssays blog, August 2026. Available online: https://stdm.github.io/Challenging-assumptions/.

  1. Prepared by Thilo & Renate Stadelmann within the COMPASS project on “Co-creating Meaningful, Positive Scenarios for Swiss Society with AI”, supported by Hasler Stiftung. ↩

  2. Scenarios that have profoundly shaped the public discourse on AI in recent years are, for example, Machines of Loving Grace, Future of Jobs Report 2025, AI 2027, The AI Index report 2026, Europe 2031, AI 2040, The Future is for Everyone, and We must build AI for people; not to be a person. Analysing them surfaces a number of assumptions that underly much of the arguments leading to the described futures, many of which holding negative consequences for much of humanity. Conducting this analysis, we experimented with up to 30 fine-grained technological, philosophical, economical, and political assumptions and ultimately aggregated them to the 8 concise assumptions treated in the remainder of this post. ↩

  3. The term “AGI” is often used interchangeably with “superintelligence”, “powerful AI”, or “strong AI” in the public discourse. ↩

Written on August 9, 2026 (last modified: August 24, 2026)